Claude docs changes for October 3rd, 2026 [diff]

Executive Summary

  • Claude Code 2.1.288 adds /code-review --max-findings, recovery of a prompt cleared with Ctrl+C, agents-view search (Ctrl+F), and many resume, auto mode, plugin and permission fixes (including a dangerous rm inside bash -c running without a prompt).
  • Self-hosted environments can now send model requests to Amazon Bedrock or Google Cloud's Agent Platform, with a new setup guide and a list of what differs from Anthropic API sessions.
  • The Claude apps gateway supports certificate client authentication (private_key_jwt) for identity providers such as Microsoft Entra, and documents a 431 troubleshooting case for users in many IdP groups.
  • UserPromptSubmit hooks are now documented as also firing on turns Claude Code starts itself (scheduled tasks, background subagent reports, cross-session messages).
  • Plugin publishing docs were reorganized into a step-by-step directory submission flow, and mods gained new pane keyboard contexts and sharing guidance.

New Claude Code versions

2.1.288

New features

  • Added --max-findings <n>|all to /code-review to report more or fewer findings than the usual limit; the choice is reused until --max-findings default
  • Added recovery of a prompt cleared with Ctrl+C: pressing Up on the empty prompt brings back the draft, including pasted text and images
  • Added Ctrl+F to find a session by name and Alt+↑/↓ to jump between groups in the agents view; both, and rename, can be rebound in keybindings.json
  • Added $.ui.selection() for mods, returning the text last selected in fullscreen mode
  • Added a built-in gh api for cloud sessions whose image has no GitHub CLI
  • Added a re-authenticate prompt when an MCP server asks for more OAuth scope during a tool call
  • Added CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS to turn structured outputs off for gateways that reject them
  • Added a screen reader announcement of the new permission mode when you approve a plan

Existing feature improvements

  • Auto mode: an overly long conversation is now compacted for the safety classifier instead of prompting for, or failing, every tool call
  • /autocompact now saves the auto-compact window per model
  • The background command time limit now applies only to unattended sessions (-p, Agent SDK, CI, cloud)
  • claude project purge is now claude purge (old name still works)
  • MCP URL prompts from servers that can't report completion now wait for "I'm done, continue"
  • Agents view: Enter on a name filter or Ctrl+F search opens the best-matching session
  • Improved Remote Control recovery from an expired server credential
  • Improved screen reader mode announcements and /permissions number selection
  • Improved Bash permission prompts with shorter reasons when part of a command can't be checked

Major bug fixes

  • Fixed a dangerous rm inside bash -c/sh -c running without a prompt in bypassPermissions mode or under a shell allow rule
  • Fixed PreToolUse and PermissionRequest hooks being skipped when matching failed; the call is now blocked
  • Fixed a BASHPID arithmetic assignment being allowed silently by the Bash permission check
  • Fixed mid-response API timeouts failing the turn in non-interactive sessions and subagents
  • Fixed long conversations failing with "Prompt is too long" instead of auto-compacting
  • Fixed several --resume problems: dropped post-compaction context, unsaved last responses, truncated transcripts, and lost thinking from 2.1.286 or earlier sessions
  • Fixed path-scoped .claude/rules and nested CLAUDE.md files not loading on Write or Edit
  • Fixed MCP tool calls sometimes running twice for oversized or unparseable results
  • Fixed /login reporting success when credentials could not be saved
  • Fixed LSP tool calls hanging indefinitely; requests now time out after 60s
  • Fixed plugin issues: LSP ${user_config.*} placeholders not substituted, git-subdir installs on older git, GitHub-source installs without an SSH key, and --plugin-dir plugins missing "Configure options"
  • Fixed idle_prompt notification hooks firing while background agents are still running
  • Fixed agent-team plugin agents running with default prompt, tools and effort

Claude Code changes

Changed documents

agent-view [Source]

  • Worktree isolation applies to sessions dispatched from agent view or started with claude --bg; a session moved to the background with ← or /background keeps editing where it was. [lines 525-551] [Source]

claude-apps-gateway-config [Source]

  • New "Certificate client authentication" section: token_endpoint_auth_method: private_key_jwt with a client_assertion block (private key and certificate) in place of a client secret, e.g. for Microsoft Entra. Includes openssl key creation, IdP upload and gateway.yaml steps; requires v2.1.284+. [lines 68-143] [Source]

claude-apps-gateway-deploy [Source]

  • New troubleshooting entry for 431 errors after sign-in when a developer belongs to many IdP groups (256 KiB header limit, limits.max_request_header_bytes). [lines 350-380] [Source]

claude-directory [Source]

  • claude project purge renamed to claude purge, with a note about the old name. [lines 1400-1451] [Source]

code-review [Source]

  • Documented the --max-findings <n>|all|default flag (also added to commands). [line 299] [Source]

env-vars [Source]

  • New CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS. [line 258] [Source]
  • New CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES limiting re-sends of timed-out non-streaming requests. [line 318] [Source]
  • MCP_PROTOCOL_NEGOTIATION now also covers stdio servers. [line 462] [Source]

errors [Source]

  • New section for API Error: Output blocked by content filtering policy: not retried; rephrase or /rewind. [lines 2697-2711] [Source]
  • Notes that the stable release channel may not reach the version a new model requires; switch to the latest channel. [lines 2399] [Source]

hooks [Source]

  • UserPromptSubmit also fires for scheduled tasks and /loop iterations, background subagent reports, and messages from other sessions. [lines 1287-1292] [Source]
  • idle_prompt is not sent while a background agent is still running. [line 2230] [Source]
  • Clarified how permissionDecisionReason is shown for ask and for denials in -p runs. [line 1764] [Source]

keybindings [Source]

  • New Pane and PaneField contexts for mod panes. [lines 63-64] [Source]
  • Reorganized the ctrl+x chord list by context, adding the pane chords (arrows, ctrl+x x) and how to unbind them. [lines 527-570] [Source]

mcp [Source]

  • Channel servers that negotiate MCP revision 2026-07-28 can't deliver channel messages; stdio servers are probed with MCP_PROTOCOL_NEGOTIATION=auto, which Anthropic is turning on by default. [line 391] [Source]

model-config [Source]

  • availableModels entries get their own /model row depending on provider. [line 250] [Source]
  • /autocompact saves the window per model, in modelSettings; the global autoCompactWindow still applies to every model. [lines 755-758] [Source]

plugins-reference [Source]

  • New LSP server requestTimeout field (default 60000 ms); same in plugins/manifest-reference. [line 332] [Source]

plugins/mods/create [Source]

  • Sharing guidance by audience: zip, own marketplace, organization-managed install, or public submission. [lines 346-352] [Source]

plugins/mods/interface [Source]

  • Pane keys: Page Up/Down, Home, End scroll; Ctrl+X arrows resize; Ctrl+X X closes. [lines 498-500] [Source]

plugins/mods/reference [Source]

  • tool.describe can set isDeferred: true to put a tool behind tool search. [line 52] [Source]

plugins/publish [Source]

  • Directory submission rewritten as steps: confirm eligibility, validate locally with claude plugin validate --strict, check what loads where, submit in the developer portal. [lines 122-148] [Source]

sandboxing [Source]

  • Removed the v2.1.199 version requirement for masking environment variables (also in settings-reference). [line 406] [Source]

self-hosted-environments-configuration [Source]

  • New "Send model requests to Bedrock or Agent Platform" section: preparing the cloud account and egress, narrowly scoped credentials and their security implications, setting CLAUDE_CODE_USE_BEDROCK/CLAUDE_CODE_USE_VERTEX, verifying the variables, and what differs from Anthropic API sessions (no server-managed settings, files, model selection, web search/fast mode). [lines 266-348] [Source]

settings-reference [Source]

  • modelSettings now accepts a per-model autoCompactWindow (100000-1000000 or "auto") alongside effortLevel and maxEffortLevel. [lines 1190-1194] [Source]
  • Per-model window takes precedence over the global autoCompactWindow. [line 2736] [Source]
  • Background session isolation setting now describes sessions moved to the background. [line 5156] [Source]

setup [Source]

  • A newly launched model can require a version newer than the stable channel; apt, dnf and apk installs choose a channel by repository. [lines 224-231] [Source]

skills [Source]

  • New precedence rule: in a local terminal session a skill replaces a built-in command of the same name, but not its aliases. [line 183] [Source]
  • Clarified how the effort frontmatter default is resolved when omitted. [line 380] [Source]

API changes

No API documentation changes today.